I have updated the error message that gets printed when trying to save a multiindex following some discussion in an ArviZ repo with @dcherian.

• [x] Related to #1077 and https://github.com/arviz-devs/arviz/issues/2165
• [ ] Tests added: haven't but can do if necessary using pytest.raises(..., match)

What is your issue?

most (if not all) of xarray documentation is written as rst files and using the ipython directive. Due to this, the html repr is not use in the documentation. I find the html repr to be much more informative and intuitive, especially for beginners and I think it would be great to use it in the documentation. There are multiple ways to do this (not necessarly incompatible between them):

• Use jupyter sphinx instead of ipython to run and embed code cells from rst files. I use this in the documentation of xarray-einstats for example
• Use jupyter notebooks instead of rst. We use this in arviz and xarray-einstats docs. However, in order to keep using all the sphinx roles and directives used the sphinx configuration would need to be modified to use myst-nb instead of nbsphinx
• Use myst notebooks instead or rst. Also used in ArviZ, also needs myst-nb instead of nbsphinx.

afaik, nbsphinx can be changed with myst-nb without needing to do any changes to the documentation, then rst files could be progressively updated to ipynb, myst or any other format supported by jupytext, rst, markdown and notebook sources can all be used at the same time to generate the documentation and link from one to the other with sphinx roles and cross references.

Is this something that sounds interesting? I could update the infrastructure at some point whenever I have time and update a page (to any or multiple of the options above) as an example and then let other people take over

Adds a mention of xarray-einstats to the ecosystem page. Related to #3322.

Modifies the code in reset_index to keep attributes when converting an indexing coordinate to non indexing coordinate. I have added tests for single index for both dataarray and dataset. Not sure if both are needed as they end up calling the same base function.

Regarding multiindex, I think it is not possible to keep the metadata as it is removed when creating the multiindex/stacking.

• [x] Closes #4101
• [x] Tests added
• [x] Passes isort -rc . && black . && mypy . && flake8
• [x] Fully documented, including whats-new.rst for all changes and api.rst for new API

If an indexing coordinate with attributes is converted to a non indexing coordinate with reset_index, the attributes are lost. I am not sure it is a bug, but I think they should keep the attributes and reset_coords does keep the attributes of reset coordinates.

MCVE Code Sample

```python temp = 15 + 8 * np.random.randn(2, 2, 3) time = xr.DataArray(pd.date_range('2014-09-06', periods=3), dims=["time"]).assign_attrs({"attr": 23}) coord = xr.DataArray([[-99.83, -99.32], [-99.79, -99.23]], dims=["x", "y"]).assign_attrs({"coord": True})

ds = xr.Dataset( { 'temperature': (['x', 'y', 'time'], temp), }, coords={ 'coord_0': coord, 'time': time, } ) ds

both coord_0 and time have attributes

ds.reset_index("time")

coordinate time_ does not have attributes anymore

ds.reset_coords("coord_0")

data variable coord_0 still has attributes

```

Expected Output

I would expect attributes to be kept.

Possible solution

I was wondering if changing this line and this other line to

vars_to_create[str(d) + "_"] = Variable(d, index, variables[d].attrs)

could solve this. If so I'll send a PR whenever I can.

Versions

This issue has nearly no effect on users, but (at least for me) it is quite bothering aesthetically. It also looks like something simple to solve even though I am not an sphinx expert so I am not sure.

In the docstring, there is a colon between the name of one argument and its type, however, in the online documentation, the name and the type are smashed together. In the case of xr.align:

I think it would be more readable and appealing if there was some kind of separation between the two.

Code Sample

I am having some troubles understanding apply_ufunc broadcasting rules. As I had some trouble understanding the docs, I am not 100% sure it is a bug, but I am quite sure. I will try to explain why with the following really simple example.

```python import xarray as xr import numpy as np

a = xr.DataArray(data=np.random.normal(size=(7, 3)), dims=["dim1", "dim2"]) c = xr.DataArray(data=np.random.normal(size=(5, 6)), dims=["dim3", "dim4"])

def func(x,y): print(x.shape) print(y.shape) return ```

The function defined always raises an error when trying to call apply_ufunc, but this is intended, as the shapes have already been printed by then, and this keeps the example as simple as possible.

Problem description

```python xr.apply_ufunc(func, a, c)

Out

(7, 3, 1, 1)

(5, 6)

```

Here, a has been kind of broadcasted, but I would expect the shapes of a and c to be the same as when calling xr.broadcast, as there are no input core dims, so all dimensions are broadcasted. However:

```python print([ary.shape for ary in xr.broadcast(a,c)])

[(7, 3, 5, 6), (7, 3, 5, 6)]

```

Using different input core dims does not get rid of the problem, instead I believe it shows some more issues:

```python xr.apply_ufunc(func, a, c, input_core_dims=[["dim1"],[]])

(3, 1, 1, 7), expected (3, 5, 6, 7)

(5, 6), expected (3, 5, 6)

xr.apply_ufunc(func, a, c, input_core_dims=[[],["dim3"]])

(7, 3, 1), expected (7, 3, 6)

(6, 5), expected (7, 3, 6, 5)

xr.apply_ufunc(func, a, c, input_core_dims=[["dim1"],["dim3"]])

(3, 1, 7), expected (3, 6, 7)

(6, 5), expected (3, 6, 5)

```

Is this current behaviour what should be expected?

Output of xr.show_versions()